The researchers behind it even mentioned that, at one point, it found ways to hop over the walls and skirt around holes, but they went back into the programming and changed it so that it wouldn’t do that.
Goddamn let him cook. Any% speedruns are the best.
It remembers me this video. An AI was trained to play a hide and seek game, and used glitches for win.
This is the only case where I would agree. And some people claim machines can’t be creative.
they hate to see a man at his best
Nice that they did this with RL but this game looks fairly straightforward to automate by more classical means.
I couldn’t find it in the article, but I skimmed the paper, and it looks like they only used a camera for input where the ball is in relation to everything.
My problem is it seems like they only used one single board configuration.
Can it do multiple varieties, with different wall spacing and such? If they get a human who’s good at these things and can do multiple layouts quickly, and a computer can do the same with randomized layouts, I’ll be more impressed.
As for whether you could automate it, sure you could with other (non optical camera) sensors to detect where the ball is at all times but doing it with just software and a camera is still mildly interesting.
If you really want to call this a physical skill game (to me, it is mostly brains and a little physical skill), then it is surely one of the most computer friendly physical skill games - two motors and you are done :-)
